SAGE: Adaptive Self-Training Towards Endogenous Memory in Large Language Models

Jiacheng Wei, Faguo Wu, DONGRUN LI, Xiao Zhang

Keywords: test-time training, self-adaptive, atomized reasoning

Abstract: Large language models (LLMs) exhibit strong generalization but face limitations in real-world adaptation, as their parameters remain static. Inspired by neuroscience and cognitive science, this work investigates endogenous memory as a mechanism for adaptive and incremental updates. We present SAGE, a framework for self-adaptive parameter updates in reasoning tasks, triggered by the detection of out-of-distribution (OOD) knowledge. SAGE consists of three core modules: (1) a Trigger module, which detects reasoning failures across multiple evaluation metrics in real time; (2) the Trigger Buffer module, which clusters reasoning failure samples using a streaming clustering process with HDBSCAN, followed by stability checks and similarity-based merging; and (3) the LoRA Store module, which dynamically optimizes parameter updates with an adapter pool for knowledge retention. Evaluation results show that SAGE demonstrates excellent accuracy, robustness, and stability on the atomic reasoning subtask through dynamic knowledge updating during test time. Specifically, an EM accuracy of $97.16\%_{\pm 4.65\%}$ reflects statistically significant and reliable performance.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 6973